Re: Re: [PATCH v7] io_uring: Statistics of the true utilization of sq threads.

From: Xiaobing Li
Date: Tue Jan 30 2024 - 02:59:45 EST


On 1/29/24 15:01, Pavel Begunkov wrote:
>On 1/29/24 07:18, Xiaobing Li wrote:
>> On 1/18/24 19:34, Jens Axboe wrote:
>>>> diff --git a/io_uring/sqpoll.h b/io_uring/sqpoll.h
>>>> index 8df37e8c9149..c14c00240443 100644
>>>> --- a/io_uring/sqpoll.h
>>>> +++ b/io_uring/sqpoll.h
>>>> @@ -16,6 +16,7 @@ struct io_sq_data {
>>>> pid_t task_pid;
>>>> pid_t task_tgid;
>>>>
>>>> + long long work_time;
>>>> unsigned long state;
>>>> struct completion exited;
>>>> };
>>>
>>> Probably just make that an u64.
>>>
>>> As Pavel mentioned, I think we really need to consider if fdinfo is the
>>> appropriate API for this. It's fine if you're running stuff directly and
>>> you're just curious, but it's a very cumbersome API in general as you
>>> need to know the pid of the task holding the ring, the fd of the ring,
>>> and then you can get it as a textual description. If this is something
>>> that is deemed useful, would it not make more sense to make it
>>> programatically available in addition, or even exclusively?
>>
>> Hi, Jens and Pavel
>> sorry for the late reply.
>>
>> I've tried some other methods, but overall, I haven't found a more suitable
>> method than fdinfo.
>
>I wouldn't mind if it's fdinfo only for now, that can be changed later
>if needed. I'm more concerned that reading fdinfo and then parsing it
>is incompatible with the word performance, which you mentioned in the
>context of using 1 vs 2 syscalls to get the stats.
>
>That can be left to be resolved later, however. Let's just be clear
>in docs that stats could be 0, which means the feature is not
>working/disabled.
>
>Another question I raised in my reply (v6 thread), why it's using
>ktime_get(), which same as jiffies but more precise, instead of a
>task time?

Sorry, I forgot to reply to you.
I was thinking wrong. you are right, we can use "getrusage" to statistics
the work_time of sqpoll in the ring.

>
>
>> If you think it is troublesome to obtain the PID, then I can provide
>
>I missed the context, where do we need to know PIDs?

Since obtaining the fdinfo content of sqpoll requires finding the corresponding
PID first, I guess Jens thinks it is troublesome to manually obtain the PID of
each sqpoll thread when there are many sqpoll threads.
Therefore, I want to write a script that can automatically output all sqpoll
statistics.

--
Xiaobing Li